Most of them are not found in ivory towers他们大多不在象牙塔里Artificial intelligence has already created trillions of dollars of market value and turned a handful of tech nerds into celebrities. The public is excited and terrified in equal measure. Many brainy people, from Bill Gates to Elon Musk, say that the technology is just getting started. Talk to academic economists, though, and most seem oddly uninterested in studying the impact of the potentially world-changing technology. The centre of gravity in AI economics is instead shifting away from universities—and a gang of “AI-pilled economists” is leading the charge.人工智能已经创造了数万亿美元的市场价值，并把少数科技宅变成了名人。公众对它既兴奋又恐惧，程度几乎相当。许多聪明人，从 Bill Gates 到 Elon Musk，都说这项技术才刚刚开始。然而，如果你去和学术经济学家交谈，会发现他们中大多数人似乎出奇地没有兴趣研究这项可能改变世界的技术会产生什么影响。AI 经济学的重心反而正在从大学转移出去，而一群“AI 入脑”的经济学家正在带头冲锋。University researchers can move fast when they want to. After Lehman Brothers, an investment bank, collapsed in 2008 and sparked the global financial crisis, economists turned the study of bank runs and credit crunches from a niche pursuit into a mainstream interest. Two months into covid-19, in 2020, close to a third of working papers on economics published by America’s National Bureau of Economic Research, NBER, a prestigious repository of economic thought, focused on the pandemic’s effects in some way. Some of this work burst into the mainstream, including by Nick Bloom of Stanford University, an expert in working from home, and Emily Oster of Brown University, who studied school closures.大学研究人员如果愿意，是可以行动很快的。2008 年，投资银行 Lehman Brothers 倒闭并引发全球金融危机后，经济学家把对银行挤兑和信贷紧缩的研究，从一个小众领域变成了主流兴趣。2020 年新冠疫情暴发两个月后，美国国家经济研究局，也就是 NBER 这个著名经济思想论文库，发表的经济学工作论文中，接近三分之一都以某种方式关注疫情影响。其中一些研究进入主流视野，比如斯坦福大学研究居家办公的专家 Nick Bloom，以及布朗大学研究学校关闭的 Emily Oster。Three and a half years after the launch of ChatGPT ushered in the AI age, by contrast, economic analysis of the technology remains comparatively scarce. The proportion of NBER papers which focus on AI is rising, but not especially fast. Even in 2024, after the covid emergency was over and the AI era had truly begun, the number of covid-related papers exceeded the number of AI-related ones. This year the NBER is likely to host more conferences on health care than on AI.相比之下，ChatGPT 发布并开启 AI 时代已经三年半了，但对这项技术的经济学分析仍然相对稀少。NBER 中关注 AI 的论文比例正在上升，但速度并不特别快。即使在 2024 年，在新冠紧急状态已经结束、AI 时代真正开始之后，与新冠相关的论文数量仍然超过与 AI 相关的论文。今年，NBER 举办的医疗保健会议数量，很可能仍会多于 AI 会议。Some academic economists have seized the AI opportunity. Susan Athey of Stanford is exploring what happens if AI puts people out of work. Basil Halperin of the University of Virginia has written lucidly about how financial markets price AI developments. Yet none is nearly as recognisable as Mr Bloom or Ms Oster. And few economists seem to recognise the research potential. “I’ve been shocked by how few of my colleagues have even tried to speak with Anthropic or OpenAI”, says one superstar academic economist who insists he talks to AI labs constantly.一些学术经济学家确实抓住了 AI 机会。斯坦福大学的 Susan Athey 正在研究，如果 AI 让人失业会发生什么。弗吉尼亚大学的 Basil Halperin 则清楚地写过金融市场如何给 AI 发展定价。然而，他们中没有人像 Bloom 或 Oster 那样广为人知。而且，似乎很少有经济学家意识到其中的研究潜力。一位坚持自己经常与 AI 实验室交流的超级明星学术经济学家说：“我很震惊，我的同事中竟然很少有人甚至尝试过和 Anthropic 或 OpenAI 交谈。”A lot of the research that exists is highly abstract. According to IDEAS/RePEc, a bibliographic database dedicated to economics, Daron Acemoglu of MIT is the highest-ranked wonk in AI economics. A paper by Mr Acemoglu published early in 2024 features a complex model of economic growth under AI, and implies modest aggregate productivity gains. It has already received more than 1,000 citations. But the model underestimates the potentially transformative effect of new AI products coming to market, argues Tyler Cowen of George Mason University. “The gains from AI measure as small because it is assumed AI will not be doing new things.”现有许多研究都非常抽象。根据专注于经济学的文献数据库 IDEAS/RePEc，麻省理工学院的 Daron Acemoglu 是 AI 经济学领域排名最高的专家。Acemoglu 在 2024 年初发表的一篇论文，建立了一个 AI 条件下经济增长的复杂模型，并暗示 AI 带来的总体生产率提升较为温和。该论文已经获得超过 1000 次引用。但乔治梅森大学的 Tyler Cowen 认为，这个模型低估了新 AI 产品进入市场后可能产生的变革性影响。他说：“AI 带来的收益被测算得很小，是因为模型假设 AI 不会做新的事情。”Many AI-linked empirical studies also appear to rest on flawed assumptions. A paper by Erik Brynjolfsson of Stanford University, ranked fifth by IDEAS/RePEc, and colleagues suggests that young people’s employment in AI-exposed occupations has sharply dropped, implying that the technology is already transforming the labour market. Yet attributing the trend to AI means believing that firms started shedding young workers upon the very first release of ChatGPT—a product that was nowhere close to being good enough to replace humans.许多与 AI 相关的实证研究似乎也建立在有缺陷的假设之上。斯坦福大学的 Erik Brynjolfsson 在 IDEAS/RePEc 排名第五，他和同事的一篇论文认为，在暴露于 AI 的职业中，年轻人的就业人数已经急剧下降，这暗示该技术已经在改变劳动力市场。然而，把这一趋势归因于 AI，就等于相信企业在 ChatGPT 最初发布时就开始裁掉年轻员工。而当时的 ChatGPT 远远没有达到能够替代人类的水平。Academic economists may be slow and sloppy for two reasons. The first relates to the type of shock that AI represents. In 2020 covid changed the world almost overnight, and the effects were visible almost instantly in the data. By contrast, AI is changing the economy under its bonnet. The average unemployment rate across the OECD, a club of rich countries, is about the same as when ChatGPT was first released. What is more, GDP numbers contain practically no AI-specific data—investment in AI data centres, for instance, can only be guessed at. With no clear macroeconomic impact and no microeconomic data, there is little for wonks to analyse.学术经济学家之所以行动缓慢而且有些粗糙，可能有两个原因。第一个原因与 AI 所代表的冲击类型有关。2020 年，新冠几乎一夜之间改变了世界，其影响几乎立刻就能在数据中看到。相比之下，AI 正在经济的“引擎盖下面”改变经济。经合组织这个富裕国家俱乐部的平均失业率，与 ChatGPT 最初发布时大致相同。更重要的是，GDP 数据中几乎没有 AI 专属数据，比如对 AI 数据中心的投资只能靠猜测。在没有明确宏观经济影响、也没有微观经济数据的情况下，经济学专家可分析的东西并不多。The second factor is that economists, as a rule, are a fairly techno-sceptical bunch. Historical research shows that technology raises incomes, but only slowly, with all sorts of non-technological factors, including financial frictions and cultural resistance, holding it back. In Britain it was decades before the technological breakthroughs of the industrial revolution translated into faster growth.第二个因素是，经济学家通常是一群相当技术怀疑主义的人。历史研究显示，技术确实会提高收入，但速度很慢，因为各种非技术因素都会拖后腿，包括金融摩擦和文化阻力。在英国，工业革命的技术突破花了几十年，才转化为更快的经济增长。A recent paper by Mr Halperin and colleagues, reporting on the results of a survey, captures this scepticism. Even under a scenario where AI progress is “rapid” by 2030—meaning that AI can compete with or surpass the brainiest humans—the median academic economist expects American GDP growth of just 3.5% in 2050, compared with 5.3% for AI researchers. Only 11% of leading economists agree that the use of AI over the next decade “will lead to a substantial increase in the unemployment rates in advanced countries”, according to a survey by the University of Chicago. If most academic economists do not think that AI will be transformational, they may prefer to stick with other research areas they consider weightier.Halperin 及其同事最近一篇报告调查结果的论文，体现了这种怀疑。即使在一种到 2030 年 AI 进展“迅速”的情景下，也就是 AI 能够与最聪明的人类竞争甚至超越他们，学术经济学家给出的中位数预期是，2050 年美国 GDP 增长率只有 3.5%；相比之下，AI 研究人员的预期是 5.3%。芝加哥大学的一项调查显示，只有 11% 的顶尖经济学家同意，未来十年 AI 的使用“将导致发达国家失业率显著上升”。如果大多数学术经济学家并不认为 AI 会带来变革，他们可能更愿意继续研究自己认为更重要的其他领域。AI-curious economists are finding a cosier home in two places away from the academy. The first is government, and in particular statistical offices and central banks. Surveys from America’s Census Bureau and Statistics Canada track AI adoption across the economy. The Bank of England’s monthly “decision-maker panel” has explored businesspeople’s perceptions of AI, while the British government recently created an “AI economics institute” to improve research on the topic. At a recent conference at the OECD, government beancounters puzzled over how to update measures of productivity for the AI age. Much of this work will not set the world on fire, but it performs a crucial public service: building the data infrastructure on which future economists will rely.对 AI 感兴趣的经济学家，正在学院之外的两个地方找到更舒适的归宿。第一个是政府，尤其是统计机构和中央银行。美国人口普查局和加拿大统计局的调查正在追踪整个经济中 AI 的采用情况。英格兰银行每月的“决策者小组”研究了商界人士对 AI 的看法，而英国政府最近建立了一个“AI 经济学研究所”，以改进这一主题的研究。在经合组织最近的一次会议上，政府统计人员正在琢磨如何为 AI 时代更新生产率衡量方法。这些工作大多不会震动世界，但它们提供了一项关键公共服务：建设未来经济学家将依赖的数据基础设施。The second, more significant place is on the front lines of the technology. In the 2010s AI labs hoovered up many brilliant computer scientists to design their models. Ufuk Akcigit of the University of Chicago and colleagues find that by 2019 more than two-thirds of AI researchers worked in industry, up from less than half in 2001. Now something similar is happening to economists.第二个，也是更重要的地方，是技术前线。2010 年代，AI 实验室吸走了许多优秀计算机科学家来设计模型。芝加哥大学的 Ufuk Akcigit 及其同事发现，到 2019 年，超过三分之二的 AI 研究人员在产业界工作，而 2001 年这一比例还不到一半。现在，类似情况正在经济学家身上发生。Anthropic has appointed Anton Korinek of the University of Virginia, who comes in second on the IDEAS ranking, to its economics-research team. OpenAI hired Ronnie Chatterji of Duke University as its chief economist. Google DeepMind, the tech giant’s in-house frontier lab, recently hired Alex Imas of the University of Chicago as its “director of AGI economics”, referring to the “artificial general intelligence” that would match or best humans at most intellectual tasks. According to The Economist’s rough tally, a few dozen AI-pilled dismal scientists have accepted jobs at the big labs.Anthropic 已任命弗吉尼亚大学的 Anton Korinek 加入其经济学研究团队，他在 IDEAS 排名中位居第二。OpenAI 聘请杜克大学的 Ronnie Chatterji 担任首席经济学家。科技巨头内部前沿实验室 Google DeepMind 最近聘请芝加哥大学的 Alex Imas 担任“AGI 经济学主任”。AGI 指的是“通用人工智能”，也就是在大多数智力任务上能够匹敌或超过人类的系统。据《经济学人》粗略统计，已有几十位“AI 入脑”的“沉闷科学家”，也就是经济学家，接受了大型 AI 实验室的工作。The attraction of an AI lab is clear. They have access to the best data, as well as the ear of policymakers. Accept a position there and before long Dwarkesh Patel, Silicon Valley’s favourite podcaster, will ask you on his show. Tech firms also have deeper pockets than universities do. Even relatively junior economist positions at an AI lab can pay $300,000 a year or more: not high relative to an AI programmer, perhaps, but well above what an early-career professor teaching Econ 101 would earn. Some lucky ones may get their hands on stock options in the world’s hottest firms.AI 实验室的吸引力很明显。它们能接触到最好的数据，也能直接影响政策制定者。接受那里的职位后，不久之后，硅谷最受欢迎的播客主持人 Dwarkesh Patel 就可能邀请你上节目。科技公司的财力也比大学雄厚。即使是 AI 实验室中相对初级的经济学家职位，年薪也可能达到 30 万美元或更多。相对于 AI 程序员来说，这或许不算高，但远高于早期职业阶段、教授经济学入门课程的大学教授。有些幸运者还可能拿到全球最热门公司的股票期权。The quality of extramural AI research is rising. In work at the Peterson Institute for International Economics, a think-tank, Mr Korinek and Patrick McKelvey of the Bank of Canada have built what they call “AI GDP” for America. The paper shows that, properly measured, it grew by more than 2,000% in both 2024 and 2025. Mr Imas publishes a useful tracker of the effect of AI on productivity. According to his judgment, there is encouraging evidence of small-scale productivity gains, but little evidence of large macro effects.校外 AI 研究的质量正在提高。在智库 Peterson Institute for International Economics 的一项研究中，Korinek 和加拿大央行的 Patrick McKelvey 为美国构建了他们所谓的“AI GDP”。论文显示，如果正确衡量，美国的 AI GDP 在 2024 年和 2025 年都增长了超过 2000%。Imas 发布了一个有用的 AI 对生产率影响追踪工具。根据他的判断，已有令人鼓舞的小规模生产率提升证据，但几乎没有大规模宏观影响的证据。All very exciting, at least to AI-pilled economic journalists. But for every clever study by Mr Korinek or Mr Imas the labs still produce a dud. Anthropic’s “economic index”, released to great fanfare, is not really an index but a random collection of data about usage of its chatbot, Claude. In March Anthropic published a report concluding that “people get better at using Claude through experience”. No duh. Last year OpenAI published descriptive work showing that 20-25% of messages on ChatGPT involved “seeking information”. Riveting stuff.这一切都很令人兴奋，至少对“AI 入脑”的经济记者来说是这样。但实验室每产出一项 Korinek 或 Imas 那样聪明的研究，也仍然会产出一些平庸之作。Anthropic 大张旗鼓发布的“经济指数”，其实并不是真正的指数，而是关于其聊天机器人 Claude 使用情况的一堆随机数据。3 月，Anthropic 发布报告，结论是“人们会通过经验变得更擅长使用 Claude”。这不是废话吗？去年，OpenAI 发布了一项描述性研究，显示 ChatGPT 消息中有 20% 至 25% 涉及“寻找信息”。真是“精彩”。No doubt the quality will improve over time. Still, if frontier AI research migrates inside firms, economists may follow the path already taken by the “tech economists” at Microsoft, Google and elsewhere. These wonks typically spend less time on big questions of social import, asking, say, whether social media is good for children, and more on narrow questions, such as how best to design auctions for selling ads. Mr Akcigit’s study notes that after making a permanent transition from academia to industry, AI researchers produce fewer papers but more patents, in effect a “reorientation from open science towards proprietary innovation”.毫无疑问，研究质量会随着时间改善。不过，如果前沿 AI 研究迁移到公司内部，经济学家可能会走上 Microsoft、Google 等公司“科技经济学家”已经走过的道路。这些专家通常较少花时间研究具有重大社会意义的问题，比如社交媒体是否对儿童有益；而更多研究狭窄问题，比如如何最好地设计广告拍卖机制。Akcigit 的研究指出，AI 研究人员从学术界永久转向产业界后，论文产出减少，但专利增加，实际上是一次“从开放科学转向专有创新的重新定位”。Then there are conflicts of interest. Lab researchers are likely to face pressure to publish work that makes AI look useful and safe. Last year Tom Cunningham, an economics researcher, left OpenAI after reportedly growing frustrated about what he could and could not publish. He ended up at METR, a research institute dedicated to evaluating AI models and the threats they pose. In a world with great possibilities but also great dangers, society needs disinterested researchers to say what they really think. Academic economists have ground to make up.然后还有利益冲突。实验室研究人员很可能面临压力，要发表让 AI 看起来有用且安全的研究。去年，据报道，经济学研究员 Tom Cunningham 因为对自己能发表什么、不能发表什么越来越感到沮丧而离开 OpenAI。他最终去了 METR，这是一家专门评估 AI 模型及其威胁的研究机构。在一个既有巨大可能性、也有巨大危险的世界里，社会需要无利益牵连的研究人员说出他们真正的想法。学术经济学家还有很多需要追赶的地方。全篇重点词汇：AI-pilled /ˌeɪ ˈaɪ pɪld/ 深度相信 AI 将改变世界的；“AI 入脑”的。strongly convinced that AI will have transformative effects on society and the economy.ivory tower /ˌaɪvəri ˈtaʊə(r)/ 象牙塔；脱离现实的学术圈。an academic or intellectual environment seen as detached from practical life.techno-sceptical /ˌteknəʊ ˈskeptɪkl/ 技术怀疑主义的。doubtful about claims that technology will quickly transform society or the economy.frontier lab /ˈfrʌntɪə læb/ 前沿实验室。a research lab working on the most advanced technologies, especially frontier AI models.conflict of interest /ˌkɒnflɪkt əv ˈɪntrəst/ 利益冲突。a situation in which someone’s personal or institutional interests may compromise their objectivity.
夜雨聆风
